Airflow's DataDriver plugin

from Pandas' dataframes to Airflow pipelines

WHY :

In a machine learning project, there is a recurring problem with the difference between local interactive modeling source code and production pipelines source code. It is very error prone and, as a consequence, time consuming because we switch constantly between experimentation and production.

The Datadriver project aims to solve this issue by making the glue code based on Pandas and sklearn for modelization, and on Airflow for automation, scheduling, and monitoring of training and predicting pipelines.

Plugin description

Datadriver UI (ddui) is the Airflow's plugin we developed to track our models. Combined with the Datadriver's API (pyddapi), it offers a DAG view to track machine learning workflow (or dataflow).

More specifically, it shows the Output of any Airflow's Task with a lot of metrics and charts :

• choose a DAG to track
  
• select a task to see charts and describe metrics on the output_table
  
• look at histograms to verify if columns are correct (distributions, number of NAs,
  unique values, etc...)
  

Getting started

from PyPI.org :

pip install ddui
ddui install # link the plugin to airflow plugin's folder    

from source install :

git clone git_url_of_this_project && cd this_project
pip install -e .
ddui install

docker install :

./run_docker.sh

Package modules

ddui/
    dash_app -> the application defined like a Dash application, with callbacks and event handeling. It is imported in plugin.py later
    dash_components -> html custom components like a Panel or an Alert Div
    orm -> function to access the Airflow metastore and retrieve DAGs list and infos
    plot -> functions using plotly, they return a Graph object
    plugin -> defines the DataDriverUI plugin that implements Airflow's Plugin interface https://airflow.apache.org/plugins.html#interface
    views -> a FlaskAdminView that implements Dash too, to have the ability to include plotly charts in Airflow

dependencies graph

Developer setup

There is an existing DAG in tests/dags that mocks the behavior of Datadriver's API, but without any dependency to pyddapi.

You can use it to develop the User Interface, using the script located in tests/dev_tools.

cd tests/dev_tools
python run_webserver.py

It runs the Airflow's webserver, and it overrides the AIRFLOW__CORE__DAGS_FOLDER to look into tests/dags.

Setup your virtual env

virtualenv venv
source venv/bin/activate
pip install -e .
pip install -r ci/tests_requirements.txt
ddui install

Contributors

This repository is a part of the DataDriver project.

Since 2016, there were many people who contributed to this project :

• Ali El Moussawi
• Arthur Baudry
• Augustin Grimprel
• Aurélien Massiot
• Benjamin Joyen-Conseil
• Constant Bridon
• Cyril Vinot
• Eric Biernat
• Jeffrey Lucas
• Nicolas Cavallo
• Nicolas Frot
• Matthieu Lagacherie
• Mehdi Houacine
• Pierre Baonla Bassom
• Rémy Frenoy
• Romain Ayres
• Samuel Rochette
• Thomas Vial
• Veltin Dupont
• Vincent Levorato
• Yannick Drant
• Yannick Schini
• Yasir Khan